Contrastive Token-level Explanations for Graph-based Rumour Detection

Wai Kit Daniel Chin, Roy Ka-Wei LeearXiv 2025 · DOI · OpenReview

TL;DR

This paper introduces CT-LRP, a novel framework that provides fine-grained, token-level explanations for Graph Neural Network (GNN)-based rumour detection models. By moving beyond node-level explanations, CT-LRP offers deeper insights into how GNNs interpret text, leading to more trustworthy and interpretable rumour detection systems.

What Problem Does This Paper Address?

Social media's rapid information spread has unfortunately accelerated the dissemination of harmful rumors. While Graph Neural Networks (GNNs) have shown great promise in automatically detecting these rumors, their "black box" nature makes it hard to understand why they make a particular prediction. Existing methods for explaining GNNs often focus on nodes (e.g., entire posts) or edges, which isn't granular enough to capture the nuances within the high-dimensional text embeddings that GNNs use as features. This lack of interpretability hinders trust and makes it difficult to debug or improve these models, especially in critical domains like public health or political discourse.

Methods

The core innovation is Contrastive Token Layerwise Relevance Propagation (CT-LRP). It builds upon existing Layerwise Relevance Propagation (LRP) techniques, which are good at tracing the contribution of inputs to an output. Here's how it works:

  1. Forward Pass: The input is tokenized text, which is then converted into high-dimensional text embeddings by a ftext function. These embeddings, along with the graph structure (adjacency matrix), are fed into a GNN-based rumour detection model (f). This produces output logits.
  2. Backward Pass:
    • GNN Explanation: Relevance scores are backpropagated through the GNN to generate node-level attributions (Rnode). This tells us which parts of the node's feature vector (embedding) were important for the prediction.
  3. Token-Level Explanation: Crucially, these node-level attributions are then backpropagated through the ftext function. This process breaks down the relevance score from the node embedding all the way down to individual tokens within the text.
  4. Contrastive Step: To ensure explanations are class-specific (i.e., relevant to the predicted rumor class and not just general text features), CT-LRP compares the token attributions for the predicted class against those for other classes. It identifies tokens that are uniquely important for the predicted class, effectively filtering out common or misleading signals. This is achieved by perturbing the input graph by removing specific token vectors and observing the change in prediction.

Datasets

The authors evaluated CT-LRP on three publicly available rumour detection datasets:

  • Twitter (combined Twitter15 and Twitter16): A four-class dataset with source posts, threads, and labels (Non-Rumour, True-Rumour, False-Rumour, Unverified).
  • Weibo: A two-class dataset similar to Twitter but based on Chinese microblogging posts.
  • PHEME: Another four-class dataset with source posts, threads, and labels.

The paper does not introduce new benchmarks but focuses on evaluating existing GNN models on these established datasets.

Key Insights and Ideas

  • Token-level is Key: The paper strongly argues that node-level explanations are insufficient for GNNs processing text. The real "action" happens at the token level, and CT-LRP successfully bridges this gap.
  • Contrastive Filtering for Specificity: The "contrastive" aspect is vital. Simply attributing relevance to tokens isn't enough; it needs to be specific to the predicted class. This step helps distinguish between tokens that are generally important for text processing and those that specifically indicate a rumor.
  • Bridging Latent Features and Actionable Insights: CT-LRP effectively translates the complex, high-dimensional latent representations learned by GNNs into understandable token-level explanations, making the models more transparent.
  • Generalizability: The framework is designed to be a post-hoc explanation method, meaning it can be applied to various pre-trained GNN models without retraining, making it broadly applicable.

Key Results

  • Superior Fidelity: CT-LRP consistently achieved the highest fidelity scores across all models and datasets, indicating that its explanations accurately reflect the model's decision-making process.
  • Improved Class-Specific Explanations: The contrastive step significantly improved the performance of token-level explanations, showing an average increase of 25.77% over a baseline token-level LRP.
  • Balanced Fidelity and Sparsity: CT-LRP achieved a superior balance between fidelity and sparsity, as measured by the Fidelity-Sparsity metric, with an average increase of 66.98% over the next best baseline. This means explanations are both accurate and concise.
  • Qualitative Example: Figure 2 provides a visual example of CT-LRP's output, highlighting tokens in blue/green for class-specific relevance and red for negative attribution, offering intuitive insights.

Limitations and Future Work

  • Implicit Graph Inductive Biases: While CT-LRP uses GNNs, it doesn't explicitly explain the graph inductive biases themselves. Future work could incorporate attribution scores that account for specific edges or subgraph structures.
  • Task and Modality Specificity: The current framework is tailored for rumour detection and text data. Extending it to multimodal tasks (e.g., text + images) or heterogeneous graphs would require modifications.
  • Future Directions: The authors suggest exploring CT-LRP's potential in other related tasks like fake news detection and misinformation challenges, as well as extending it to more complex real-world scenarios involving multimodal and heterogeneous data.

Summary updated

Abstract

The widespread use of social media has accelerated the dissemination of information, but it has also facilitated the spread of harmful rumours, which can disrupt economies, influence political outcomes, and exacerbate public health crises, such as the COVID-19 pandemic. While Graph Neural Network (GNN)-based approaches have shown significant promise in automated rumour detection, they often lack transparency, making their predictions difficult to interpret. Existing graph explainability techniques fall short in addressing the unique challenges posed by the dependencies among feature dimensions in high-dimensional text embeddings used in GNN-based models. In this paper, we introduce Contrastive Token Layerwise Relevance Propagation (CT-LRP), a novel framework designed to enhance the explainability of GNN-based rumour detection. CT-LRP extends current graph explainability methods by providing token-level explanations that offer greater granularity and interpretability. We evaluate the effectiveness of CT-LRP across multiple GNN models trained on three publicly available rumour detection datasets, demonstrating that it consistently produces high-fidelity, meaningful explanations, paving the way for more robust and trustworthy rumour detection systems.
BibTeX
@misc{chin2025contrastive,
  author = {Wai Kit Daniel Chin and Roy Ka-Wei Lee},
  title = {{Contrastive Token-level Explanations for Graph-based Rumour Detection}},
  year = {2025},
  eprint = {2502.04366},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.04366},
}
Figure